基于深度卷积判别网络的人脸比对方法  

Face-matching method using deep convolution discrimination network

作　　者：谷凤伟 陆军[1,2] 刘子玄 蔡成涛 GU Fengwei;LU Jun;LIU Zixuan;CAI Chengtao(College of Intelligent Systems Science and Engineering,Harbin Engineering University,Harbin 150001,China;Key Laboratory of Intelligent Technology and Application of Marine Equipment(Harbin Engineering University),Ministry of Education,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学智能科学与工程学院,黑龙江哈尔滨150001 [2]哈尔滨工程大学船海装备智能化技术与应用教育部重点实验室,黑龙江哈尔滨150001

出　　处：《哈尔滨工程大学学报》2024年第9期1770-1782,共13页Journal of Harbin Engineering University

基　　金：国家自然科学基金项目(52171332);黑龙江省自然科学基金项目(F201123).

摘　　要：针对实际应用中人脸比对面临着场景复杂性高、光照、遮挡等问题,为了提高人脸比对准确率,本文提出了一种基于深度卷积判别网络的人脸比对算法MTC-FaceNetSDM。建立了MTC-FaceNetSDM的深度卷积神经网络,在FaceNet网络前端中融合多任务级联卷积神经网络得到MTC-FaceNet网络,实现实际场景中的人脸检测提取目标人脸;利用深度卷积神经网络获取高维人脸深度特征,并将FaceNet网络的欧氏距离模块替换为所提出的相似度判别模块SDM,用于高维人脸特征向量比对;最终,利用自制的人脸数据集C-facev1,结合CASIA-WebFace人脸数据集对本文人脸比对算法进行训练,使用人脸数据集LFW和CASIA-FaceV5对训练后的模型进行性能评估。实验结果表明:本文所设计的MTC-FaceNetSDM的人脸比对准确率比MTC-FaceNet整体提高1.48%,对中国人脸比对准确率提高3.80%,可实现多人种的人脸比对,同时该算法具备良好的鲁棒性和泛化能力,达到优良的人脸比对效果,可实际应用于人脸验证系统。Aiming at the problems of high scene complexity,illumination,and occlusion in face matching in practi-cal applications,this paper proposes the face-matching algorithm MTC-FaceNetSDM based on a deep convolution discrimination network to improve the accuracy of face matching.First,the deep convolutional neural network framework in MTC-FaceNetSDM was established,and the MTC-FaceNet network was obtained by embedding a mul-titask cascaded convolutional neural network in the front of the FaceNet network structure.Then,the deep convolu-tional neural network was used to obtain high-dimensional face depth features,and the Euclidean distance module in the FaceNet network structure was replaced with the proposed similarity discrimination module(SDM)for high-dimensional face feature vector matching.Finally,the self-made face datasets C-facev1 and CASIA-WebFace were used to train the face-matching algorithm proposed in this paper,and the face datasets LFW and CASIA-FaceV5 were used to evaluate the performance of the trained model.The experimental results showed that the face-matching accuracy of MTC-FaceNetSDM was 1.48%higher than that of MTC-FaceNet.Moreover,the Chinese face-matching accuracy was increased by 3.80%,thus showing the proposed algorithm′s capability for multiethnic face matching.Moreover,the proposed algorithm had favorable robustness and generalization ability,achieving excellent face com-parison results,which could be practically applied to face verification systems.

关 键 词：人脸比对 深度卷积判别网络 多任务级联卷积神经网络 相似度判别模块 人脸特征向量 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]